基于机器视觉算法和改进MobileNetV2的晶圆缺陷检测研究  

Research on Wafer Defect Detection Based on Machine Vision Algorithms and Improved MobileNetV2

作　　者：赖肖强 胡泓 LAI Xiaoqiang;HU Hong(Shenzhen Key Laboratory of Advanced Manufacturing,School of Mechanical and Electrical Engineering and Automation,Harbin Institute of Technology(Shenzhen),Shenzhen 518063,China)

机构地区：[1]哈尔滨工业大学(深圳)机电工程与自动化学院深圳市先进制造重点实验室,广东深圳518063

出　　处：《机械与电子》2025年第1期34-39,46,共7页Machinery & Electronics

基　　金：深圳市科技计划资助项目(JSGG20201201100410029)。

摘　　要：基于机器视觉以及AOI自动光学检测平台,设计了一套针对晶圆表面缺陷的自动化检测系统。首先设计了晶圆缺陷检测的总体方案,分别对AOI系统的机械结构以及运动控制模块制定合理方案。然后基于机器视觉算法设计晶圆缺陷检测的算法,用于检测晶圆表面的光敏面脏污、光敏面划痕和电极环断开缺陷。同时基于改进后的MobileNetV2模型实现对晶圆表面电极脖子区域的缺陷检测。这一套方法可以实现晶粒的缺陷检测,脏污、划痕、电极环断开等缺陷的检测准确率分别达到94.06%、97.37%、92.36%,平均单颗晶粒检测耗时分别为86 ms、13 ms、93 ms,使用基于改进MobileNetV2模型的电极脖子缺陷检测准确率达到99.3%。以上实验结果表明,所提方法准确率高、耗时短,可以满足晶圆缺陷检测的需求。This paper introduces an automated defect detection system for wafer surfaces,combining machine vision and AOI technology.The overall approach includes designing mechanical and motion control modules for the AOI system,implementing machine vision algorithms to detect various wafer surface defects,and employing an enhanced MobileNetV2 model for electrode neck defect detection.The system achieves high accuracy rates of 94.06%,97.37%,and 92.36%for detecting contamination,scratches,and electrode ring disconnection,respectively.The average detection times per wafer grain are 86 ms,13 ms,and 93 ms for these defects.Additionally,the electrode neck defect detection accuracy using the improved MobileNetV2 model reaches 99.3%.These results demonstrate the proposed method’s effectiveness in achieving accurate and efficient wafer defect detection.

关 键 词：晶圆 缺陷检测 机器视觉 AOI 图像处理 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN307[自动化与计算机技术—计算机科学与技术]